PROJECT SUMMARY- TR&D2: DIFFUSION MRI AND CONNECTOMICS In our new TR&D2 project, Diffusion Imaging & Connectomics, we extend our decades of work in the mathematics of diffusion MRI to create new and useful tools. Our work draws on our last 4 years of productive collaborations, which led to over 100 papers. We focus on 3 tasks: (1) modeling white matter microstructure, (2) automatically extracting maps of fiber bundles, and (3) modeling brain connectivity in more powerful and adaptive ways. Each method seeks to overcome a major problem that affects the validity of diffusion MRI today. In Aim 1, we develop mathematical metrics of white matter microstructure that can better use the available information in low-quality dMRI, but drastically boost the analytical power of higher-quality dMRI. With our diverse range of Collaborative Projects - including one that collects ultra-high resolution diffusion spectrum imaging at high-field (7T DSI - we test extensions of our tensor distribution function model of dMRI, which outperforms the standard tensor model and the most popular metric, DTI-FA. We extend the TDF to bi-exponential and multi- compartmental models, using TV-L1 fusion to merge all metrics across the image to better detect disease. In Aim 2, we extend our tools for automatic tract labeling. Our new approach, FiberNET, uses deep learning to avoid the manual intervention required by our initial tool, autoMATE. We will compare it head to head for accuracy, power, and utility on all our Collaborative Project datasets. These include diffusion MRI data from HIV+ children, from elderly people with various types of dementia, and adolescents at risk for psychiatric disorders. In Aim 3, we use two approaches - L1 fusion (dFUSE) and connectivity based on continuous functions (ConCon) - to overcome 2 very serious problems with the analysis of brain networks - the problem of large number of false positive fibers in tractography, and the arbitrariness of picking a cortical parcellation. These problems affect all downstream network metrics, to the point where the networks are very hard to compare across studies, with current methods. ConCon models the connectivity directly as a density, making it amenable to registration and statistics without defining specific borders in the cortex; dFUSE combines the reliable fibers across many tractography methods, boosting the contribution of methods that are more reliable on a given dataset. All our tools will be widely disseminated. With our broad collaborative network, we hope to advance the neuroscience of brain connectivity by using novel mathematics, tested head to head against existing methods.